Making Serverless Pay-For-Use a Reality with Leopard

Explore a conference presentation that challenges current serverless computing billing practices and introduces an innovative solution for true pay-for-use pricing. Discover how researchers from the University of Wisconsin-Madison analyzed existing serverless billing models and found that current practices fail to align with actual resource consumption, particularly under the prevailing SLIM (static, linear, interactive-only model) assumptions. Learn about the Nearly Pay-for-Use (NPFU) billing model, a new approach that accommodates varying CPU and memory demands, spot cores, and preemptible memory to achieve more accurate resource-based pricing. Examine the design and implementation of Leopard, an NPFU-based serverless platform that integrates billing awareness across multiple critical subsystems including the CPU scheduler, OOM killer, admission controller, and cluster scheduler. Understand how this billing-aware architecture addresses the limitations of traditional serverless platforms and enables more efficient resource utilization. Review experimental results demonstrating Leopard's benefits for both cloud providers and users, including throughput increases of more than 2x and significant cost reductions, making serverless computing more economically viable and resource-efficient for realistic workloads.